Previously, we would rely on the later phases to error out
(often for using too much shared memory). This enables the
checks on the IR that already exist for CUDA and OpenCL also
for ROCm.